AI agents are getting real — not because the demos are prettier, but because teams are finally getting serious about the plumbing: speed, governance, and repeatable workflows.
Here are a few companies and techniques worth watching right now, with a practical builder’s lens.
1) FloQast: visual agent builders (no-code, but still real workflows)
Accounting is one of those industries where “close enough” is not acceptable. FloQast’s new Visual Agent Builder is interesting because it targets repetitive financial operations work and tries to make agent creation accessible with drag-and-drop (AI Agent Store — March 2026 news).
Technique to steal: treat agents like a workflow with checkpoints. If a step impacts money, inventory, or compliance, add a “human sign-off” gate and clear logging. Agents shouldn’t be magic; they should be measurable.
2) Qualys: security agents that don’t just find issues — they fix them
Qualys launched Agent Val, described as an agent that identifies security problems, confirms which threats are real, and then closes the gaps automatically (AI Agent Store — March 2026 news). That “confirm, then act” loop is the pattern to copy.
Technique to steal: use a two-step agent loop:
- Detect: gather evidence and produce a short, structured report (what’s broken, where, and why it matters).
- Remediate: propose a patch or change, then run validation (tests, linting, policy checks) before it touches production.
3) Google Cloud: the unsexy but critical work — inference efficiency
If you’re building agents, token costs and latency will hit you fast. Google Cloud laid out five practical techniques to push LLM systems closer to an “efficient frontier”: semantic routing across model tiers, prefill/decode disaggregation, quantization, context routing + prefix caching, and speculative decoding (Google Cloud — efficient frontier of LLM inference).
The concrete takeaway: routing and caching can deliver big wins without changing models. In Google’s case study, optimizing routing produced 35% faster time-to-first-token on Qwen3-Coder workloads and doubled prefix cache hit rate from 35% to 70% (Google Cloud — efficient frontier of LLM inference).
Technique to steal: if your product uses a big system prompt or the same RAG “starter context” repeatedly, implement prefix caching before you spend a dollar on bigger GPUs.
4) Multi-token prediction: making one model generate faster (no extra draft model)
A recent multi-token prediction approach aims to convert standard next-token models into parallel decoders by training them to emit spans of tokens at once, with confidence-adaptive decoding at runtime (InfoWorld — multi-token prediction). The reported results: 3×+ speedups on GSM8K with <3% accuracy loss for an 8B model (InfoWorld — multi-token prediction).
Technique to steal: if you’re operating at scale, don’t just choose a model — choose a decoding strategy. Sometimes the cheapest improvement is a better generation loop.
The shop-floor takeaway
Agents are moving from “chat” to “work,” but the winners won’t be the ones with the fanciest prompts. They’ll be the ones who can answer three questions:
- Speed: How do we keep latency down when the agent runs 20 steps?
- Safety: What gates stop it from doing something expensive or dumb?
- Repeatability: Can we log it, test it, and run it the same way next week?
Get those right, and agentic systems stop being a novelty and start being a competitive advantage.